Meta has introduced Muse Spark, its latest artificial intelligence model and the first to come out of its newly formed Superintelligence Labs. The model is designed as a next-generation large language system with improved reasoning capabilities, including a dual-mode setup that balances speed and deeper, multi-step thinking. Unlike Meta’s earlier open-source efforts, Muse Spark is being rolled out as a more controlled, product-focused technology, built to integrate directly across platforms like Facebook, Instagram, and WhatsApp.
Muse Spark is positioned as the first model in a new ‘Muse’ family (internally referred to as ‘Avocado’), which the company describes as a stepwise scaling approach where each generation validates the architecture before expanding to more powerful systems. Technically, the model is designed to be relatively small and efficient, yet capable of handling complex reasoning tasks across domains like science, mathematics, and healthcare.
In internal benchmarks, it has shown competitive performance against leading models from OpenAI, Google, and Anthropic, though it does not consistently outperform them across all categories. For example, in high-level reasoning tests like GPQA Diamond, it trails the top-performing frontier systems but remains within a comparable range.
A defining feature of Muse Spark is its dual-mode reasoning architecture. In its standard mode, the system delivers rapid responses similar to conventional chat-based AI. However, users can switch to a more advanced ‘thinking’ or ‘contemplating’ mode, where the model deploys multiple sub-agents that work in parallel to break down complex problems. This multi-agent coordination allows the system to tackle tasks like planning, analysis, or multi-step decision-making with greater depth. Meta reports that this approach significantly improves performance on difficult benchmarks.
The model is also multimodal, meaning it can process both text and images. This allows use cases like analyzing photos, extracting contextual insights, and generating responses based on visual input. Healthcare is a key focus in Muse Spark’s development, with input from over 1,000 doctors to improve the accuracy of medical and wellness responses. The model can provide clear, structured answers on health topics, though concerns about reliability remain. It is also designed for everyday use cases like coding help, education, travel planning, and general decision-making.
Another major focus is deep integration with Meta’s existing ecosystem. Muse Spark currently powers the Meta AI assistant in the company’s standalone app and website, with plans to expand across WhatsApp, Instagram, Facebook, Messenger, and AI-enabled devices. With access to more than 3.5 billion users globally, Meta aims to embed AI directly into social interactions, content discovery, and communication.
From a product and business perspective, Muse Spark also introduces new monetization opportunities. The model includes features that transform social content into commerce experiences, like recommending products based on posts, creators, and brand activity within Meta’s platforms.
Despite its capabilities, Muse Spark is not being released as an open-source model. Access is currently limited to Meta’s own platforms and a restricted API preview for selected partners. Meanwhile, the latest launch reflects a major shift in Meta’s AI strategy after a year of restructuring and heavy investment. In 2025, it set up Superintelligence Labs, bringing together about 3,000 employees and multiple AI teams, backed by over $14 billion and the hiring of AI leader Alexandr Wang. Importantly, the move also followed internal dissatisfaction with earlier models like LLaMA 4. The development becomes even more significant as the Mark Zuckerberg-led firm recently revealed it could spend up to $135 billion on AI infrastructure alone in 2026.